wi-go: accurate and scalable vehicle positioning using wifi fine...

Wi-Go: Accurate and Scalable Vehicle Positioning using WiFiFine Timing Measurement

Mohamed IbrahimWINLAB, Rutgers

[email protected]

Ali RostamiWINLAB, Rutgers

[email protected]

Bo YuGeneral Motors Research

[email protected]

Hansi LiuWINLAB, Rutgers

[email protected]

Minitha JawaharWINLAB, Rutgers

[email protected]

Viet NguyenWINLAB, Rutgers

[email protected]

Marco GruteserWINLAB, Rutgers

[email protected]

Fan BaiGeneral Motors Research

[email protected]

Richard HowardWINLAB, Rutgers

[email protected]

ABSTRACTDriver assistance and vehicular automation would greatly bene-fit from uninterrupted lane-level vehicle positioning, especially inchallenging environments like metropolitan cities. In this paper, weexplore whether the WiFi Fine Time Measurement (FTM) protocol,with its robust, accurate ranging capability, can complement cur-rent GPS and odometry systems to achieve lane-level positioning inurban canyons. We introduce Wi-Go, a system that simultaneouslytracks vehicles and maps WiFi access point positions by coherentlyfusingWiFi FTMs, GPS, and vehicle odometry information together.Wi-Go also adaptively controls the FTMmessaging rate from clientsto prevent high bandwidth usage and congestion, while maximiz-ing the tracking accuracy. Wi-Go achieves lane-level vehicle posi-tioning (1.3 m median and 2.9 m 90-percentile error), an order ofmagnitude improvement over vehicle built-in GPS, through vehicleexperiments in the urban canyons of Manhattan, New York City,as well as in suburban areas (0.8 m median and 3.2 m 90-percentileerror).

CCS CONCEPTS• Networks→ Location based services.

ACM Reference Format:Mohamed Ibrahim, Ali Rostami, Bo Yu, Hansi Liu, Minitha Jawahar, VietNguyen, Marco Gruteser, Fan Bai, and Richard Howard. 2020. Wi-Go: Accu-rate and Scalable Vehicle Positioning using WiFi Fine Timing Measurement.In The 18th Annual International Conference on Mobile Systems, Applications,and Services (MobiSys ’20), June 15–19, 2020, Toronto, ON, Canada. ACM,New York, NY, USA, 13 pages. https://doi.org/10.1145/3386901.3388944

Permission to make digital or hard copies of all or part of this work for personal orclassroom use is granted without fee provided that copies are not made or distributedfor profit or commercial advantage and that copies bear this notice and the full citationon the first page. Copyrights for components of this work owned by others than ACMmust be honored. Abstracting with credit is permitted. To copy otherwise, or republish,to post on servers or to redistribute to lists, requires prior specific permission and/or afee. Request permissions from [email protected] ’20, June 15–19, 2020, Toronto, ON, Canada© 2020 Association for Computing Machinery.ACM ISBN 978-1-4503-7954-0/20/06. . . $15.00https://doi.org/10.1145/3386901.3388944

1 INTRODUCTIONThe rapid evolution of advanced driver assistance and vehicle au-tomation systems, along with their growing market [1, 2], haveled to increased demand for lane-level vehicle positioning that isaccurate even in urban canyon environments. Example applicationsfor such solutions include lane-level navigation and vehicle safetycommunications.

Today’s vehicles primarily use the Global Positioning System(GPS), often in conjunction with vehicle odometry for correctingshort termGPS biases. However, inmany challenging environments,such as urban canyons, bridges and tunnels, multi-path fading orshadowing considerably degrades satellite positioning accuracy.While research has shown how position accuracy can be furtherimproved to a fewmeters withmotion sensors andmapmatching [9,17, 20, 34] in some urban environments, these still face challengesin more extreme urban canyons. To achieve lane-level positioning,highly instrumented automated vehicle prototypes use cameras orLiDAR sensors to reference their measurements against availabledetailed models and imagery of the roadway. Creating, maintaining,and making available such detailed image models for all roadwaysis laborious and resource demanding, since it can undergo frequentchanges due to reasons such as snow or falling leaves.

To reduce reliance on such resource intensive image registration,we investigate whether WiFi time-of-flight ranging as specified inthe WiFi Fine Time Measurements (FTM) standard [8] are suffi-ciently uncorrelated with GPS measurements to achieve lane-levelpositioning in urban canyons. While WiFi [19, 58] positioning or isfrequently used in smartphones, these have been based on receivedsignal strength (RSS) positioning which is limited to an accuracyof tens of meters in urban canyons. In contrast, FTM ranging canachieve meter-level accuracy in open-space environments [3] [31].Given this promise of improved accuracy and its wide availability,our study focuses onWiFi FTM and explores its utility in a vehicularcontext.

To address the lane-level urban canyon positioning challenge,this paper introduces Wi-Go1, a scalable and accurate vehicle track-ing technology that complements GPS and odometry with time-of-flight measurements and multilateration to surrounding access

1Wi-Go stands for WiFi, GPS and odometry based tracking

https://doi.org/10.1145/3386901.3388944

https://doi.org/10.1145/3386901.3388944

MobiSys ’20, June 15–19, 2020, Toronto, ON, Canada Ibrahim, et al.

points through the recently standardized WiFi Fine Time Measure-ment (FTMs) protocol. It can opportunistically use WiFi accesspoints deployed in buildings or in cars parked along the roadway.Since the access point location is usually not known a priori, wedesign FTMSLAM, a collaborative simultaneous localization andmapping approach to simultaneously track landmark (APs andparked vehicles) positions as well as moving vehicle positions. Re-alizing this requires addressing several key issues.

First, moving from a passive-client approach as in conventionalRSSI-based WiFi positioning, to the actively-signaling-client ap-proach used in Time-of-Flight based tracking, can introduce con-tention and channel congestion, particularly in densely populatedareas such as urban canyons. This creates challenges when scalingto larger numbers of vehicles and access points. Consider a scenarioin which hundreds of nearby vehicles send periodic FTM requestson the same WiFi channels. Since every FTM request triggers asequence of packet transmissions specific to one client, this caneasily exhaust the available channel capacity, introduce latency andtherefore degrade the positioning accuracy for these clients. Notethat this was not a concern for RSSI-based positioning since manyclients can passively overhear the same access point message butdue to the lack of synchronized clocks FTM requires round trippacket exchanges. To tackle this issue, Wi-Go mitigates channelcongestion by optimizing the FTMmessage rates to each AP to max-imize the tracking accuracy, while remaining below a cumulativeallowed message rate in a given region.

Second, the FTM ranging process involving a series of packetexchanges introduces time offsets between the range measurementsto individual access points, while standard multilateration methodin wireless localization assume quasi-simultaneous ranging to mul-tiple landmarks. This is particularly important in the vehicularcontext where a fast moving vehicle can travel a significant dis-tance between individual measurements. To address this challenge,we propose Mobile Multilateration, a novel tracking algorithm, totrack a moving vehicle with individual ranges obtained at differentpositions, while still leveraging odometery information to relateeach position of the vehicle to its previous position.

Third, to minimize receiver complexity, it requires simultaneouslocalization and mapping (SLAM) techniques that only use rangeinformation while existing SLAM work heavily relies on bearingmeasurements as well (e.g.,[30, 51]). Wi-Go presents a range-onlySLAM framework suitable for multilateration with Fine Time Mea-surements; it explicitly represents the uncertainty of both the posi-tion estimates via particle filters, and can use FTM measurementsfrom a second antenna to resolve AP position ambiguity on linearroadways.

Fourth, in environments such as urban canyons, WiFi communi-cation can also be heavily affected by multipath fading and shad-owing. Simply using WiFi measurements that are potentially lowquality will not yield the desired accuracy improvements over GPS.Therefore we design an uncertainty weighted multilateration tech-nique that estimates and explicitly considers the quality of thecurrent GPS and FTM measurements when updating the vehicleand access point location.

We prototyped Wi-Go with Intel 8260 wireless cards on a smallform factor PC installed with roof-mounted antennas in a vehi-cle and implemented backend algorithms in a cloud server. We

(a) WiFi Vs GPS (b) Vehicle linear movement

Figure 1: (a) WiFi edge over GPS. (b) Effect of linear move-ment of a vehicle on AP localization error.

evaluated Wi-Go in two deep urban canyon environments of up-per Manhattan as well as in a suburban residential setting. Wi-Goachieves 1.3 m median error for vehicle tracking, with relativelylow AP density (4 APs), in our urban canyon dataset in which abaseline of a built-in GPS only achieves 9.04 m median localizationerror. Meanwhile, Wi-Go’s adaptive algorithm can efficiently adaptthe FTM message rate while still showing 41.7% improvement interms of vehicle localization error.

Summary of Contributions. Wi-Go and its FTMSLAM algo-rithm make the following contributions compared to earlier posi-tioning and SLAM algorithms:

• Designing a novel FTM-SLAM algorithm to intelligentlycomplement GPS and odomentry with range-only WiFi FTMdata, while addressing the resulting latency and multi-pathchallenges.

• Mitigating channel congestion due to active time-of-flightranging messages, by intelligently distributing the cumula-tive allowed message rate over the nearby APs.

• Introducing the opportunistic use of parked vehicles forpositioning by incorporating pairwise-distance estimatesbetween APs in parked vehicles.

• Demonstrating and validating that theWi-Go systemwith itsuse of WiFi FTMs meets lane-level positioning requirementsthrough extensive experiments in two challenging urbancanyon environments and a suburban setting, with a smallfleet of 5 research vehicles.

2 BACKGROUNDIn this section, we will review the WiFi Fine Timing Measurement,SLAM, and vehicle sensing concepts that this work builds on. WiFipositioning has evolved from initial fingerprinting [10] approachesto using Angle-of-Arrival [59], dead reckoning [57] and Time-of-Flight (ToF) [25]. The ToF-based WiFi distance measurements haverecently been incorporated in WiFi standards, and promise wide-spread AP deployment support.

WiFi Fine Timing Measurement (FTM). IEEE 802.11-2016Standards [8] has included the Fine Timing Measurement protocol,802.11REVmc, to perform wireless ranging by measuring the roundtrip time (RTT) between an AP and a WiFi station (STA). The pro-tocol subtracts processing times from the round trip time, convertsit into a one-way time-of-flight estimate, and uses this to estimate

Wi-Go: Accurate and Scalable Vehicle Positioning using WiFi Fine Timing Measurement MobiSys ’20, June 15–19, 2020, Toronto, ON, Canada

range using typical propagation speed. As an interactive protocol, itconducts multiple message exchanges to achieve a higher rangingaccuracy. Recent research [31] has confirmed that the FTM protocolcan achieve meter-level accuracy in open space environments butthe accuracy degrades in high multipath environments.

Android [4] and major vendors of WiFi chipsets have started tosupport the FTM protocol. However, their accessibility to PHY layerinformation like Channel State Information is quite limited [28]since the station has to associate with the AP, which takes severalseconds [46]. This is not suitable for fast moving vehicles.

SLAM. Classic SLAM frameworks like FastSLAM [43, 51, 53]require ranging technology (e.g. LiDAR, or stereo cameras) thatestimates the distance and bearing to a landmark. Range-only SLAMwith RF or sonar beacons[21, 22, 27, 44, 45], approximates the vectorconnecting current pose with current location estimate of landmark,requiring the initial landmark location estimate to be accurate.Previous range-only SLAM frameworks used in robotics [16, 22,45] has initialized landmarks through taking majority votes overmultiple solutions. Each of the solutions either fits a pair of ranges(assuming the robot does not move in a straight line) or drawsprobabilistic samples from a circle around the robot location withradius equal to the range. These two initialization approaches mayconverge to the actual location and achieve high accuracy if therobot is not moving in a straight line, which is not the case in ourapplication (as shown in Fig. 1(b)).

On the other hand, current WiFi SLAM [15, 23, 24, 30] eithertreats RSSI fingerprints as landmarks and tracks them simulta-neously with user’s location, or augments WiFi with cameras toestimate bearing to the RSSI-based fingerprints. None of these priorworks have estimated the APs’ locations simultaneously whiletracking a user, because current propagation models that map RSSIto distance are not accurate enough and cannot generalize to differ-ent environments. WiGo introduces a new SLAM approach calledFTMSLAM, to track APs and vehicles with range-only FTM rangemeasurements. FTMSLAM uses novel opportunistic sensing algo-rithms including vehicle to AP bearing estimation, adaptive FTMrange calibration learning, and vehicle location correction throughAPs street maps.

On-board vehicle sensors.Modern vehicles are equipped withsensors to measure vehicle speed, steering wheel angle, heading,yaw rate, etc. While this information is usually communicated onthe CAN bus (and often readable from the On-Board Diagnostics(OBD-II) port), its encoding is specific to certain vehicles and pro-prietary [33]. In this paper, we leverage available on-board sensorsto feed our FTMSLAM framework with precise odometry. More-over, we leverage this odometry information as an independent,orthogonal source for correcting FTMs in over- and under-estimatecases.

3 CANWIFI AUGMENT GPS?Our Wi-Go system design is motivated by the observations thaturban canyons, a key environment with degraded GPS, usuallyenjoy a dense deployment of WiFi Access Points (APs) (as shown inFig. 1(a)). Current WiFi outdoor localization[32, 36] mainly countson RSSI fingerprinting, which is shown to be less accurate due toseveral innate limitations, like multipath effects [10], variation of

(a) Urban Experiment

0 1 2 3 4 5 6 7

Ranging error (m)

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

CD

F

FTM

(b) CDF

Figure 2: WiFi FTM ranging accuracy.

Approach Vehicle GPS GPS Android Loc.Error (m) 9.04 18.2 19.6

Table 1: Median tracking error of related technologies inManhattan, NY.

performance between different devices of the same vendor [40],and low granularity [29].The recently available WiFi Fine TimeMeasurements, instead, promise a more robust, accurate, and fine-grained ranging measurement to enable mobile multilaterationtechniques, suggesting a possibility that WiFi FTM ranging couldfurther complement GPS.

To examine our hypothesis, we conduct a preliminary experi-ment in an urban canyon environment (Manhattan, NYC). In thisexperiment, we place 4 access points on top of 4 parked vehicles as‘virtual’ WiFi APs, and evaluate the ranging and location error, inthe rest of the paper, using the following four different technologyoptions:

(1)WiFi FTM baseline. Our WiFi AP is equipped with an IntelDual Band Wireless-AC 8260 chipset that supports FTM capability;(2) Standalone GPS. Our vehicle is equipped with an after-marketU-blox EVK-7P GPS receiver (< 1m precision in an open sky setup);(3) Vehicle GPS. Meanwhile, our vehicle retrieves vehicle GPSreadings from OBD-II port. This built-in GPS has been internallycorrected using the vehicle IMU and odometry sensors; (4)AndroidFusion Location API. Finally, we collect location measurementsthrough Android Fused Location Provider API from a Google Pixel3a smartphone that is placed on the vehicle’s dashboard. The An-droid Fused Location Provider API fuses multiple sensors includingGPS, WiFi RSSI, and cell tower positioning together to providelocalization information. These technologies summarize the state-of-the-art outdoor localization approaches which are either GPS,motion sensors dead-reckoning, or based on WiFi RSSI, CellularRSSI, or a fusion of some of them.

In order to acquire ground truth locations of the vehicle, twoIntel RealSense Depth Cameras are mounted on both sides of thevehicle’s roof that log depth information of both sides of the road.We examine the recorded frames looking for landmarks such aslight poles and trees and use these landmarks’ positions to infer thevehicle’s position. Specifically, when the landmark appears in themiddle of the field of view, we use the landmark’s location alongwith its depth to compute the current vehicle location.


Figure 3: Wi-Go System Architecture Overview.

Figure 4: Illustration of our Mo-bile Multilateration. A vehiclemoves from point a, to b, thenfrom b to c.

Fig. 2(b) shows the ranging error of only WiFi FTM as othercompeting technology options do not report range measurements.The ranging error of WiFi FTM is 0.95 m median error (and 2.9 m at90-percentile CDF curve). On the other hand, we also analyzed thelocalization errors of vehicle GPS, Android Fused Location, and stan-dalone GPS as shown in Table 1. This preliminary result indicatesthat WiFi FTM seems to hold great promises of complementing andaugmenting existing outdoor localization techniques.

In the next sections, we describe in detail the design of Wi-Go,which achieves this promise of lane-level positioning.

4 THEWi-Go SYSTEMThe objective of Wi-Go is to provide a practical positioning systemthat scales in real world scenarios where tens of APs and hundredsof vehicles compete in the shared WiFi channels through activeWiFi FTM ranging. Besides scalability, we aim for uninterruptedmeter-level positioning by augmenting WiFi FTMs with GPS andodometry readings.

4.1 System OverviewIncorporating WiFi FTMs for outdoor positioning is challengingdue to the following reasons: First, with extended FTM ranginglatency, a fast vehicle could register multiple locations at differentnearby APs, imposing a vehicle tracking challenge; second, FTMsare affected by multipath and are a range-only measurement thatdoes not offer bearing information, which makes jointly locatingAPs and vehicles challenging; third, the injection of FTM packetsfrom many vehicles onto the same WiFi channels causes networkcongestion;

To address the above challenges, we design Wi-Go as shownin Fig. 3. In each participating vehicle, Wi-Go collects WiFi FTMsfrom the surrounding APs, along with the timestamp, GPS read-ing, and the vehicle’s speed and heading through on-board sensors.We assume that vehicles will at least occasionally enter open-skyGPS conditions, and that this could be used to establish positionin the world coordinate frame. The Wi-Go system then tracks avehicle by starting with an initial location estimate, gradually re-fining based on successive GPS readings, vehicle odometry, andWiFi FTM range measurements to access points positions. It uses

simultaneous localization and mapping techniques to jointly es-timate vehicle position and access point positions. As in currentWiFi positioning systems, access point position estimates can beshared across vehicles through a server.

WiGo addresses the fast moving vehicle challenge throughmobile multilateration, compensates for multipath throughuncertainty-aware fusion, and controls FTM request through acongestion-aware optimization. We describe these techniques inthe following subsections.

4.2 Mobile MultilaterationTracking a rapid-moving vehicle is not an easy task, since it requiresquick ranging due to vehicle speed. If a vehicle moves at 20 m/s andthe standard FTM latency is 0.2 s, then the next ranging measure-ment will be at least 4 meters away. A new form of multilaterationis thus needed for a moving vehicle to collect sequential rangingmeasurements at different precise locations from multiple nearbyAPs.

Fig. 4 illustrates this issue: A vehicle starts ranging at locationA and obtains a ranging measurement r1 from AP1, then it gets r2fromAP2 at location B, and finally it collects r3 fromAP3 at locationC. Our goal is to find the current location of the vehicle v, given APslocations [x1, x2, ..., xn ] and corresponding ranges [r1, r2, ...,rn ]. Totackle this issue and relate the measurements of location A and Bto location C, we leverage vehicle odometry readings to estimatedisplacement and heading between these different locations. For-mally, we formulate this problem of Mobile Multilateration as anon-linear least squares formulation:

argminvn

n∑i(| |xi − vi | | − ri )

2

subject to vi = reckon(vi−1,di−1, θi−1 + 180),(where i = 1, ...,n)

(1)

Through this formulation, we could identify the current locationof the vehicle by minimizing the least square error while back-wardly holding the constraint functions. For instance, with the lastdisplacement and heading, we can derive the current location vnfrom the preceding location vn−1.


Figure 5: FTM Weighting. (a) Triangular inequality. (b) As-signing measurement confidence based on a time window.

4.3 Uncertainty WeightingDuring the collection of measurements, weights are assigned toFTMs and GPS readings, representing the confidence of these read-ings, which will be used later in the fusion as described in Section6.

FTM Weighting. We use vehicle wheel encoders and inertialsensors to assign weights to FTM measurements. In this step, wedo not correct collected range measurements; instead, we simplyassign weights to these measurements and feed them along withtheir weights to the FTMSLAM algorithm (Section 6). Triangle in-equalities are used to evaluate measurement confidence: in Fig. 5(a),let d be the estimated displacement distance reported by the inertialsensors between two FTM measurement locations. Also, let r1 andr2 be the measured FTM ranges at these two locations. Then the fol-lowing triangular inequality must be satisfied: |r1−r2 | ≤ d ≤ r1+r2.For each FTMmeasurement r(t), we evaluate the triangle inequalityof r(t) with another FTM measurement within a small time window(Fig. 5(b)). The weight wFTM (t) for measurement r (t) is thus theratio of measurement pairs that satisfy the triangle inequality overthe total number of pairs.

GPS Weighting. GPS NMEA (National Marine Electronics As-sociation) sentences are used to obtain data which is leveragedto assign a weight for the GPS location. For a given time t , thisdata includes: signal-to-noise ratios SNR(t) for each observed satel-lite, number of observed satellites n(t), GPS quality qGPS (t), andhorizontal dilution of precision HDoP(t). GPS quality indicates3 different states: no fix (qGPS (t) = 0), GPS fix (qGPS (t) = 1),or differential GPS fix (qGPS (t) = 2). HDoP measures the geo-metric quality of GPS satellites configuration in the sky [37]. Thesmaller the HDoP number, the better the geometry in terms ofbeing spread in the sky, which has been reflected into the GPSlocation precision. The GPS weight for the system is calculated aswGPS (t) =

qGPS (t )∗n(t )∗mean(SNR(t ))HDoP (t ) .

4.4 Congestion-Aware Adaptation of FTMRequest Transmission

When many vehicles try estimating their positions using requestsand responses from multiple APs nearby, WiFi channel congestionmay occur. Since FTM is an active measurement, it is initiated fromthe STA and consumes bandwidth for every burst it sends out. Withhundreds of vehicles and tens of APs sharing limited WiFi channels,it is critical to limit how many FTM requests a vehicle can send outunder fixed bandwidth restriction. This challenge is addressed inthis section by adapting the samples per burst parameter (spb).

Problem Formulation. Thanks to pioneering studies on con-gestion control in vehicular networks [12–14], we assume that thevehicle estimates the maximum message rate limit that it is allowed

to send without causing congestion. In this paper, the main taskis thus to distribute and provide message rates for each vehicle todifferent APs nearby, while still honoring the cumulative messagerate limit and minimizing the vehicle location error. Note that themessaging between vehicles and APs consumes channel bandwidthand that potentially affects APs as well. To address this, specificchannels or partial bandwidth could be reserved for FTMmessagingor a bandwidth limit can be set for such messaging. This will inturn determine the maximum message rate per vehicle.

To achieve this goal, as shown in the objective function (Eq. 3),we aim to minimize the key factors that contribute to vehicle local-ization error collectively: the geometry of the chosen subset of APs(1 − δ (spb)), error model in AP location (eAP ), and FTM rangingerror er (reflected in the error covariance matrix C∆r ).

Approach. A naive solution would be to distribute rate limitequally over surrounding APs; however, not all APs in range canprecisely estimate a vehicle’s location. Therefore, it is importantto determine the subset of nearby APs the FTM requests shouldbe sent to. Collinearly distributed APs, for instance, fail to providean accurate position estimation; similarly, APs that are in distantlocations provide range measurements with larger errors.

Motivated by the above observations, we instead use a weightedround robin schedule for ranging to each nearby AP; in particular,we intentionally adapt the samples per burst parameter for each APon one hand, while maximizing the vehicle localization accuracyon the other hand. For an AP, the spb can be set to zero if this APis unlikely to be useful, and the sample rate adapts based on thelocation of a vehicle and its AP counterpart, respectively.

Given that x is the AP’s location estimates, v(t − 1) is thelast location estimate for the vehicle, and r(t) is the vector ofrange measurements to these APs, the ranging model would ber(t) = distance(x, v(t)) + e. e represents the total error originatedfrom FTM ranging errors er plus errors in AP’s locations (eAP ).As distance is not a linear function of locations of APs and a ve-hicle, we can linearize these equations using initial estimates ofx and v(t). Through this linearization, we can determine correc-tions to the estimates and obtain current location of the vehicle:∆r = A∆v + e, where A is an nx3 matrix of the partial derivatives(Jacobian) of the distance function with respect to the unknownvehicle locations. Through a standard least squares solution (∆v),the covariance matrix of the solution is:

C∆v = (ATC−1∆rA)

−1 (2)

whereC∆r represents the error contributed by eachAP. As this erroris dependant on the location of AP (e.g., some APs could be moreaffected by multipath than others), these errors cannot be assumedto be the same across different APs. We therefore only assume theerrors of APs are independent, and the standard deviation of thetotal error for each AP can be calculated as σ =

√σ 2eAP + σ

2er .

To optimize the vehicle localization error, we derive our objectivefunction from the covariance matrix of the least squares solution(Eq. 2) as follows:

argminspb

((1 − δ (spb))ATC−1∆rA)

−1

subject to∑

spbi = ratel imit

spbi ≥ 0, 1 ≤ i ≤ |APs |

(3)


where δ (spb) has the same form as the Delta function and eAP, errepresents the error of APs’ location estimations and range mea-surements respectively. In this model, prior AP location estimatesare assumed to exist already and each estimate has its error derivedfrom the covariance of that estimate; at the same time, newly ob-served APs are be ranged occasionally through a separate processin order to establish an initial position estimate.

The error model used for FTM range estimates can be estimatedempirically by fitting the linear regression function (eri = a ×

disti + b) using some real world data [31]. The ranging error isapproximated using the Central Limit Theorem when spb samplesare taken and averaged. For large spb, the variance of that error can

be approximated as σ 2

spbi: σer =

√(a×distance(x,v(t−1))+b)2+σ 2

empspbi

.So the ranging error to a certain AP is estimated based on its lastestimated distance and the empirically obtained standard deviationσemp of these measurements.

5 THE FTMSLAM ALGORITHMAll of these ideas are put together in the FTMSLAM algorithmicframework, the cornerstone of theWi-Go system. In FTMSLAM, thevehicle and the surrounding WiFi APs are localized and tracked, byincorporatingWiFi FTMs, GPS, and on-board sensor measurements.

Novelty. FTMSLAM utilizes range-only measurements (FTM),in direct contrast to conventional SLAM approaches which requireboth range and bearing measurements. To conquer this challenge,we intentionally take advantage of a suite of novel methods, ve-hicular opportunistic sensing, to update and correct the locationsof vehicle and APs respectively. The opportunistic sensing meth-ods include (1) vehicle to AP bearing estimation (Section 6.2), (2)resampling particles from FTM multilateration to refine estimatedlocation (Section 6.3), and (3) adaptive FTM range correction (Sec-tion 6.3).

5.1 Location Modeling and InitializationLocation Modeling. Inspired by classic FastSLAM [53], we modelthe vehicle location using particle filters, which provides a non-parametric probabilistic position representation. Through thismodel, we update a vehicle’s particles through dead reckoning,and then further correct the distribution of particles using the fusedlocation estimate of GPS andWiFi FTMmeasurements. At the sametime, we model the AP location as a Gaussian distribution. In ourstudy, it is found that the lack of angular information in FTM mea-surements introduces several new challenges: the requirement foraccurate initialization, the need for fusing GPS and FTM measure-ments for the correction step, and the demand to estimate a rangecalibration factor for different APs. These issues are addressed inSec 6.2.

APLocation Initialization. This is done through pre-collected,crowd-sourced GPS traces along with FTMs. We estimate the lo-cations of APs using uncertainty-weighted mobile multilateration(Section 4). To be specific, the optimization problem is formulatedas follows: given the location vi (t) of vehicle i at timestamp t , thecollected FTM range ri j (t), and their weightswFTM

ij (t), the goal isto identify the location of unknown AP x j . To derive the accuratelocation information, we compensate the length of wired cables

(a) (b)

Figure 6: Angle estimation illustration.

used to extend WiFi antennas by subtracting c from all the FTMranges ri j (t). We thus solve this mobile multilateration problemthrough the following weighted non-linear least squares formula:

argminx j

Nv∑i

T∑twFTMij (t)(| |xj − v(t)| | − (ri j (t) − c))2 (4)

5.2 Opportunistic Bearing EstimationUsing dual antennas connected to a WiFi transceiver placed onthe vehicle, the bearing is estimated from the vehicle to an AP. Asillustrated in Fig. 6, two antennas are fixed on the right and left edgesof a vehicle to obtain differential measurements across the directionperpendicular to the vehicle bearing. In practice, nonetheless, thenoise in FTM readings could exceed the actual difference betweentwo measurements (left antenna and right antenna), which is highlydepending on the width of the vehicle. Even worse, FTM protocolscurrently implemented on commercial WiFi transceivers do notgrant access to any PHY-layer information that could help estimatethe angle of arrival (e.g. signal phase).

To overcome this practical challenge, we consider the fact that wecan estimate which side the APs are placed as well as the minimumrange rmin to that AP over the first round of FTM measurements.Later, when we have another estimated minimum range througha new visit to this AP, we take the weighted average between thenew estimate and the historical estimates to improve estimationaccuracy. In the subsequent rounds, we can estimate the bearinginformation to that AP, whenever we reach the nearest point tothe AP (at this point, the bearing should be either 90 deg or 270 degfrom the vehicle’s heading). As the vehicle moves forward, wecan estimate the distance d from the nearest point to the AP byintegrating the current vehicle’s speed. With rmin and d , we canestimate θ as illustrated in Fig. 6(a): θi = tan−1 rmin

di. This equation

can be generalized for any road shape as illustrated in Fig. 6(b).Given rmin , d , and ϕ, we can estimate θ : θi = tan−1 x+rmin

y − ϕ,x = sin(ϕ) ∗ d , y = cos(ϕ) ∗ d .

In this algorithm, we heavily rely on the minimum range to theAP, which is attributed to the fact that FTMs are more accurateover short distances. As a result, the estimated bearing throughthe algorithm is relatively accurate; however, the accuracy coulddecrease as d increases, since the accumulation of speed sensorerror negatively affects the estimated d . To estimate the distancefrom the vehicle to the AP more accurately, we combine the twoFTM measurements from the right and left antennas respectivelythrough a weighted average using the FTM weights.


5.3 FTMSLAM Tracking FrameworkFTMSLAM Initialization Step. FTMSLAM requires initializingboth the vehicle’s location estimate and the APs’ location estimatesfirst. To do so, we use a GPS measurement with a sufficiently largeGPS weight, or alternatively, through WiFi FTM multilaterationwith sufficiently accurate estimated locations of surrounding APs.

Update/Prediction Step. In this step, we update the locationsof the vehicle and the currently discovered APs. To update a ve-hicle’s location, we estimate its displacement from its previouslocation, together with the vehicle’s heading using its in-vehiclesensors (these sensors provide vehicle kinematic information suchas speed, steering wheel angle, heading, and yaw rate). Here, the dis-placement can be estimated as follows: d = 1

2 (st −st−1)∆t . Throughestimating both displacement and azimuth, which is the headingwith respect to the north, we can update the vehicle’s location us-ing Vincenty’s formula[55] which takes the Earth curvature intoaccount.

In the meantime we also update the locations of surroundingAPs, through Extended Kalman Filter (EKF) [53]. Given that thebearing to APs is not always available, we follow a range-onlySLAM approach for opportunistically updating APs if the bearingestimation becomes unavailable. Here the location of the AP isupdated, by using the vector linking the current estimate of ve-hicle’s location with previous estimate of this AP location. Sincethis vector depends on the AP location initialization, an accurateinitialization is crucial for such a range-only approach. To furtherimprove the accuracy, the bearing estimate is improved when thebearing estimation algorithm is invoked.

Correction Step. In this step, we update the weight of eachparticle. By correcting the particle’s distribution of the vehicleusing both GPS and FTMs, we update the location of the vehicleand avoid the accumulation of motion sensors error. There are avariety of different places: in rural or suburban areas where theGPS is accurate but WiFi is not available, and the reverse couldbe true in urban canyons where GPS reception is poor, but denseAPs have already been deployed. It is likely that we can correct thedistribution of particles based on fused correction weight betweenGPS and FTM estimate.

With dense deployment of APs, we can estimate the currentvehicle’s location using weighted multilateration by using FTMs,their weights (wFTM

j (t)) and current location estimates of theseAPs. Based on the distance between each particle and the currentFTM estimate for the vehicle’s location, the FTM weight of thatparticle is updated,wFTM

par ti (t). However, this dense enough deploy-ment of APs is not always available. In this case, the weight of theparticle is updated based on the average of the error between rangemeasurements and estimated range, which is the distance betweenthe particle and the current location estimate of each AP. Similarly,the GPS weight of that particlewGPS

par ti (t) is updated based on thedistance between current GPS estimate and the particle.

To fuse these two weights, we take the weighted average of thesetwo particle weights:wpar ti (t) = w

FTM (t) ∗wFTMpar ti (t)+w

GPS (t) ∗

wGPSpar ti (t). As the FTM weight of that particle is estimated by lever-

aging FTM measurements to multiple APs, we estimate this as theaverage of these FTM weights, so wFTM (t) =

∑Nj wFTM

j (t)/N ,where N is the number of currently seen APs.

FTM Resampling. In urban canyons, a vehicle exhibits de-graded GPS readings, leading to the particles being tens of metersaway from its actual location. This requires an aggressive and fastway of re-weighting the particles when the vehicle starts to ob-serve a sufficient number of APs for multilateration. In such cases,when the current estimated vehicle location through FTM is faraway from the last estimated location, all particles will end up withzero weights. To resolve this particle deprivation problem, we re-sample a small portion of the particles from the FTMmultilaterationlocation.

Adaptive FTMRange Correction. FTMweighting cannot mit-igate multipath effects by itself, as multipath error in certain areascan be consistent and ubiquitous. An adaptive mechanism for cor-recting FTM range is thus needed to mitigate multipath effects.

In each update step, FTMSLAM uses the error between the es-timated range (FTM) and the corresponding distance between anAP and a vehicle to update AP location. After multiple iterations,the accuracy of AP location estimate is seen to improve, leadingto a better estimation of vehicle location. Inspired by this observa-tion, we could calculate a range calibration factor to compensatemultipath error, since the distance error to an AP would eventuallyconverge to its actual range error. In practice, a 2D spatial mapof the range calibration factor for each AP is built by clusteringlocations with similar patterns of multipath effects. Hence similarcorrection values and estimated calibration factors could be usedto compensate similar multipath effects.

6 PARKED VEHICLES AS PSEUDO APSIn addition to opportunistically using stationary WiFi access pointsin buildings or roadside infrastructure that support the WiFi FTMstandard, Wi-Go can be extended to also take advantage of WiFidevices in parked vehicles as Pseudo WiFi APs. As a vehicle parks,it can switch from WiFi station (STA) mode to AP mode throughrelatively straightforwardWiFi firmware changes, if vehicle batterymanagement system permits. Using parked vehicles as pseudo APscould effectively increase the density of WiFi APs, and also addhigh-quality pseudo APs since parked vehicles at the street curb arelikely to have a line-of-sight propagation to moving vehicles. Thisinnovative idea, nonetheless, will change several design choices weoutlined in Section 4-6. In this section, we intend to briefly discussthese issues and also shed light on their high-level solutions.

Pseudo AP Location Initialization. The last estimated loca-tion of a parked vehicle (just before the ignition switches off) couldserve as an initial location for that pseudo AP. More importantly,these parked APs can estimate distance to surrounding APs, beforeswitching to a AP mode. This method can effectively estimate thepairwise distances between a subset of APs, which helps jointlyinitializing APs locations. We designed a different optimizationmethod for parked, stationary vehicles because parked vehicles canobtain pair-wise rang measurements between them. Specifically,we can solve the multilateration problem jointly for all NAP APs,instead of solving the problem independently for each AP:

argminx

NAP∑j

Nv∑i

T∑tw FTMi j (t )( | |xj − v(t ) | | − (ri j (t ) − c))2

subject to | |xj − xk | | = r jk , k = 1, 2, ..., NAP

(5)


0 2.5 5 7.5 10 12.5 15 17.5 20 22.5 25 27.5 30

Localization error (m)

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

CD

F

GPS

GPS+Odometry

Smartphone Fused Location

Wi-Go

(a) Parked Cars (114th st.)

0 2.5 5 7.5 10 12.5 15 17.5 20 22.5 25 27.5 30

Localization error (m)

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

CD

F

GPS

GPS+Odometry

Smartphone Fused Location

Wi-Go

(b) Indoor APs (49th st.)

Figure 7: Vehicle tracking error of Wi-Go Vs current local-ization systems in Manhattan experiments.

Approach Parked Cars Indoor APsWi-Go 1.9 m 3.6 mGPS+Odometry 14.3 m 16.4 mGPS 18.9 m 27.4 mSmartphone Fused Loc. 17.8 m 16 m

Table 2: AP localization mean error in Manhattan experi-ments.

Pseudo AP LocationModeling. Parked vehicles bring anotherchallenge regarding modeling of pseudo AP location: Regular APsinside of buildings rarely change locations (maybe once a year);now, with these parked vehicles, pseudo APs could move morefrequently (multiple times per day). To tackle it, we model pseudoAP location probabilistically using a particle filter approach thattakes vehicle mobility into account.

Vehicle Battery Duty Cycle. Powering an AP while a vehicleis parked, may raise a battery issue if the pseudo APs are kepton for a long duration. To tackle that, we could either develop anintelligent power management algorithm that relies on BluetoothLow Energy to wake up WiFi AP unit, or adopt a simple solutionto stop vehicles from switching to Pseudo APs if its battery is lessthan a pre-set threshold (e.g. 80%).

The design and implementation of this idea merit an independentstudy, and we leave it for future work.

7 EVALUATION7.1 Experimental SetupWiFi FTMs. We setup our vehicle with a small form factor com-puter that contains two Intel Dual Band Wireless-AC 8260. EachWiFi transceiver connects to two WiFi external antennas: 6dBi RP-SMA Dual Band 2.4GHz 5GHz with 1.637 m cable to attach theantennas on the roof (we subtract that length, c from FTM ranges).We leverage an open Linux FTM tool [31] to initiate and extractFTMs from theseWiFi cards. On the AP side, we use ASUSWireless-AC1300 RT-ACRH13 APs which are configured to respond to FTMrequests as a built-in capability.

Vehicle Odometry. We connect OBDLink MX to the vehicleOBD-II interface, and retrieve odometry readings to the smallform factor computer through Bluetooth interface (Plugable USBadapter).

7.2 Urban Canyon EvaluationIn this section, we evaluate Wi-Go in terms of localization errorthrough an urban canyon experiment in which GPS is significantly

Source Single FTM ProcessingMedian Latency (ms) 20 2.2

Table 3: Wi-Go Latency.

degraded. Moreover, we show how Wi-Go will react in terms oflatency, as many APs and vehicles are actively contending overWiFi channels through WiFi FTM packets and usual WiFi datatraffic in urban canyon.

Experiment Summary. Fig.2(a) illustrates one of our experi-ments in upper Manhattan, New York City. We park four vehicles ina single street (182.6m long) as shown in the figure as red stars, andwe place an AP on each vehicle. In the fifth vehicle, we use our formfactor PC configured as a WiFi station (STA), which continuouslyranges surrounding APs.We also collect odometry, vehicle GPS, andstandalone GPS readings from this vehicle. The fifth vehicle drivesdown the street multiple times to gain statistically meaningful re-sults. We repeated the same experiment in Midtown Manhattan,where we placed six APs inside local shops.

Ground Truth. In obstructed environments like urban canyons,even high precision GPS will experience large errors. It is alsoinfeasible to manually measure a vehicle’s location while driving.Due to these reasons, getting accurate ground truth to compare ourresults to was challenging. We obtain ground truth by leveragingdepth cameras as mentioned earlier in Sec. 3. To validate our groundtruth methodology, we compare depth readings of a light pole fromdepth camera and laser range finder (or measuring tape) over arange of 3 to 12 m in a parking lot environment. The resulting erroris within 0.73 m. To obtain the ground truth location of the APs,we take pictures of the APs’ surrounding environments (buildingsand other landmarks), cross reference with Google Street View, anddrop pins accordingly on Google Maps to obtain the coordinates ofthe APs. Prior studies show that Google Earth has an accuracy closeto 1 m in metropolitan cities like Montreal [26] and Rome [47].

For these experiments, we compare the performance of Wi-Goto other technologies mentioned earlier. In Fig. 7(a), we show thecumulative distribution function (CDF) of the localization error ofthe vehicle using Wi-Go. Our Wi-Go achieves 1.3 m median error,and 2.8 m 90-percentile. In contrast, vehicle GPS (GPS+Odometry)achieves 9.04 m median error, and 13.5 m 90-percentile, comparedto standalone GPS achieving 19.6 m median error, and 42.1 m 90-percentile. Finally, Android Fused Location Provider API achieves18.2 m median error, and 58.09 m 90-percentile. In another experi-ment (Fig. 7(b)), where APs were inside of local shops, the mediantracking error of Wi-Go is 2.1m, while 90-percentile error increasesto 6.5 m due to extra signal degradation caused by concrete wallsof these shops.

Table 2 shows the AP localization error of Wi-Go system whencompared to baseline technology options. As the vehicle movesacross the street with more rounds of driving, Wi-Go improves theAP localization accuracy, reaching 1.9 m mean localization erroracross all the APs. This result demonstrates a significant perfor-mance improvement of Wi-Go system over baseline solution. Forthe baseline achieving 14.3 m mean error, we utilize the traces viavehicle GPS to estimate the locations of APs using multilaterationand FTM ranges.


Median latency is reported in Table 3. The median latency toacquire a single FTM reading, i.e. the time difference between themoment that we make a system call to initiate FTM process to themoment that call returns with range measurements from a singleAP. We measure this median latency, in our current setup, to be 20ms. On the other hand, the median processing latency of FTMSLAMis 2.2 ms. The total latency of Wi-Go system is controlled by thelatency of extracting the sensing information such asWiFi FTM andOBD readings. With a normal vehicle speed, this measured latencymay lead to error of a few meters. We believe that this latencycould be improved in the actual production system by optimizingthe FTM extraction tool.

Simulating Dense Environment.We use ns-3.30.1 to simulatea scenario of dense environment including 900 vehicles and 200APs. The mobility trace is generated using SUMO mobility simu-lator [5] for an approximately 500 m radius around Times Square,NYC, imported from OpenStreetMap [6]. We imported a map of theselected neighborhood in Manhattan to create both the mobilitytraces of the vehicles and determine the location of buildings toincrease simulation realism.

In this simulation, we implemented the WiFi FTM protocol fol-lowing the technical specification of IEEE 802.11 standard [8] (e.g.,packet format and packets size). For system parameters that are notdirectly regulated in the standard (e.g., expiration timers and othersupporting parameters), we directly infer them from our empiricalexperiments. We log the latency for getting a single FTM rangingin our simulation.

Simulation Parameter ValueSimulation time 30 secTransmission power 16.5 dBmChannel bandwidth 80 MHzChannel number 155Line-of-sight reference pathloss 21.87 dBLine-of-sight pathloss exponent 3.39

Table 4: Simulation configuration.

Table 4 shows the configuration used for the simulation. To bettercapture the impact of the multipath effects in an urban environment,the simulator distinguishes obstructed, none-line-of-sight (NLOS)communications from line-of-sight (LOS) communication as fol-lows: when there is a packet transmission, the simulator identifiesthe communication category and then chooses an appropriate prop-agation model to calculate the received signal at the receiver. Theobstructed building signal propagation model is an implementationof Mangel et al. [41]. For the line-of-sight model we used pathlosscomponents from a recent industry consortium (CAMP) [7] empir-ical measurement study, which considers the impact of vehiculartraffic condition on the pathloss.

Fig. 8(a) illustrates the median latency of different approaches,grouped together for specific vehicle densities. The error bars indi-cate 25th% and 75th% of the latency for each bars. Vehicle density iscaptured by the percentage of cars equipped with Wi-Fi, which wecall vehicle penetration ratio, and simulation results are presentedfor four penetration ratios. It is clearly shown that Wi-Go, labeledas Adaptive spb, significantly outperforms the other competingapproaches.

(a) (b)

Figure 8: (a) 25th%, median, and 75th% latency for Wi-Go(adaptive spb) compared to baselines using ns-3 simulations,and (b) mean and standard deviation of the latency versusnumber of completed AP rangings in one AP list scan for75% penetration ratio.

(a) Residential Experiment

0 1 2 3 4 5 6

Distance to precise GPS (m)

00.10.20.30.40.50.60.70.80.9

1

CD

F

Wi-Go

(b) CDF

Figure 9: Evaluating vehicle tracking accuracy in the resi-dential apartment complex experiment.

Approach Mean Min MaxWi-Go 2.6 m 1.3 m 4 mGPS+Odometry 3.6 m 2.9 m 4.3 m

Table 5: AP localization error in the residential apartmentcomplex experiment.

Fig. 8(b) illustrates the end-to-end latency from the moment thata vehicle starts with a list of APs until it successfully finishes its en-tire FTM session with the nth AP, averaged across entire simulation.Note that the vehicle starts over with a fresh list of scanned APseither after finishing the current list of APs or after 2.5 s, whichevercomes first. The shaded area around each curve shows the standarddeviation. It is shown that on average the vehicle can start triangu-lation by having the third FTM session completed after 500 msecwith our proposed adaptive spb approach; in contrast, using 10spband 20spb approaches, the vehicle has to wait approximately 1500msec and cannot complete ranging with more than seven APs.

7.3 Residential Apartments EvaluationSummary. In our second experiment, we evaluate Wi-Go in aresidential apartment complex area. We place three APs insidestudent apartments, in the location where residents usually keeptheir access point, as shown in Fig. 9(a). The choice of apartmentunits was limited to a set of volunteers who provided access to theirunits. Our data collection has been scheduled over a week with two


trips each day (morning and afternoon). One ‘trip’ refers to eachmarked colored path in Fig. 9(a) (e.g. blue path is one trip).

Ground truth. Ground truth locations of APs are estimated byfinding the latitude and longitude of the nearest window on GoogleMaps and correcting with manually measured AP distance to thewindow. Regarding vehicle localization error, we use high precisionGPS as ground truth to evaluate vehicle localization error.

Fig. 9(b) shows the CDF of the vehicle localization error for thisresidential area. Wi-Go, with vehicle initialization through accurateGPS measurement, achieves 0.8 m median localization error, and3.2 m 90-percentile. Table 5 summarizes the AP localization errorfor Wi-Go compared to the baseline solution (GPS corrected withodometry readings). Our Wi-Go can determine the locations of theAPs with 2.6 m mean error over all APs compared to 3.6 m meanerror for the baseline. For this baseline result, we filtered noisy GPSreadings using our uncertainty weights for GPS, and then used FTMmeasurements and multilateration to determine the AP’s position.

7.4 Micro BenchmarksEffect of varying number of APs. The density of APs is an im-portant factor affecting vehicle localization error. In Fig. 10, westudy this effect in the Manhattan experiment, in which WiFi FTMdominantly affect localization error compared to suburban areas.This figure shows that, as expected, the median of vehicle localiza-tion error decreases with increasing number of APs, as we increasemore reference points.

Effect of different FTMSLAM algorithms.We compare herethe impact of our proposed algorithms over standard multilater-ation on AP localization error. Fig. 11 shows that our algorithmscan gradually improve the average localization error of APs. Inour urban canyon dataset with low-quality GPS measurements,adding our GPS and FTM weights to the standard multilaterationdoes not improve the accuracy significantly. In contrast, as we ap-ply FTMSLAM, our tracking framework with our crafted weights,the average localization error drops to 5.8m. When we apply ourcomplete FTMSLAM, by correcting vehicle’s particles with FTMmultilateration estimate and bearing estimation, the AP localiza-tion error is further reduced to 1.9 m. In the residential experiment,Fig. 11 shows that our GPS and FTM weights could significantlyimprove the average localization error of APs, since a number ofoutliers exist in these measurements. After we apply completeFTMSLAM, the AP localization error decreases from 4.2 m to 2.6m.

Effect of varying FTM correction factor. FTM ranging is af-fected by multipath as shown in previous work [31]. We study howthe correction factor, which we subtract from the FTM ranges, couldaffect FTMSLAM in terms of AP localization error. Fig. 12 indicatesthat there is an optimization value that minimizes the AP local-ization error. According to our evaluation, this value is differentbetween our two datasets; as expected, it is higher in the residentialdataset compared to the urban canyon dataset. This is because, inthe residential scenarios, the APs are placed inside the buildingand thus more susceptible to multipath, in direct contrast to urbanscenarios where APs are on the top of the vehicles characterizedwith Line-of-Sight signal propagation. Our FTMSLAM derives this

FTM correction factor automatically through our adaptive FTMrange correction algorithm.

AP Location Convergence.Wi-Go improves APs location es-timations over time. Wi-Go can converge to meter-level accuracyof an AP location estimation with relatively small number of visitsto this AP. Specifically, Wi-Go can converge in Manhattan to 1.9m, 3.6 m average AP localization error after 10, 3 visits to parkedcars, indoor APs (114th st., 49th st.) with 488, 3273 average FTMreadings, respectively. The number of readings is a function of thenumber of visits and the traffic condition (average vehicle speed).Our system can also converge to 2.6 m AP localization error after14 visits to a residential area with 150 average FTM readings. Notethat as more vehicles visit an AP, AP location accuracy improvesand, as a result, vehicle localization accuracy improves.

7.5 Evaluating Adaptive FTM RateIn this section, we evaluate our solution to adapting the FTM mes-sage rate to maximize vehicle localization accuracy, while beingconstrained to a cumulative message rate.

Experimental Results. We conduct an experiment in a subur-ban area, in which we place 12 APs (first seven are indoors and therest are outdoor) as shown in Fig. 13(a). In this experiment we aimto show the importance of wisely ranging to a subset of the nearbyAPs, as well as to study how this mechanism affects the vehiclelocalization accuracy. To do so, 12 APs are deployed in an area of117m X 36.5m, and we assume that the locations of these APs arealready estimated through previous steps (wardriving multilatera-tion, or FTMSLAM with minimal sample per burst). To avoid anycomplication, we do not use any preprocessing algorithms (e.g.,FTM uncertainty weighting or FTMSLAM), but we utilize nearbyAPs to track the vehicle using standard multilateration and countonly onWiFi FTMs. We use a high precision GPS receiver as groundtruth to calculate the final localization error.

Fig. 13(b) compares our adaptive approach to the default ap-proach using fixed spb per AP (2 and 10). As expected, the 2spbapproach is susceptible to FTM noise, resulting in a high local-ization error of 9.2 m median error, and more importantly, 50 m90-percentile. On the other hand, by increasing the parameter to10spb, it is expected to improve the location accuracy, as there aremore samples to average out the noise. However, to our surprise,this higher spb approach generates 11.5mmedian error, and 52.1m90-percentile. This counter-intuitive observation could be explainedas follows: Though we increased the sampling set for statisticalaccuracy, we also significantly increase the opportunity to collectmore inaccurate, noisy data from faraway APs.

Our proposed adaptive approach, instead, takes all these factorsinto account. Our adaptive algorithm achieves 6.7 m median error,and 24.6 m 90-percentile, respectively. These results validate thatour approach improves the localization error, since it intelligentlyselects the subset of APs which are nearby, less affected by FTMnoise, and with a lower horizontal dilution of precision.

7.6 DiscussionIn this section we discuss the major lessons learnt from our studyand the results obtained from different scenarios are related andscrutinized.


1 2 3 4

Numer of APs

0

1

2

3

4

5

6

7

Ve

hic

le lo

ca

liza

tio

n e

rro

r (m

ed

ian

(m

))

Figure 10: APs density effect onvehicle’s localization error inManhattan.

Urban canyon Residential0

5

10

15

20

25

30

AP

lo

ca

liza

tio

n e

rro

r (a

ve

rag

e(m

))

MultilaterationWeighted Multilat.FTMSLAMFTMSLAM (with multilat. correction, bearing estimation)

Figure 11: FTMSLAM algorithms ef-fect on AP localization error.

0 1 2 3 4 5 6 7 8 9 10

Ranging correction parameter (m)

2

3

4

5

6

7

8

AP

lo

ca

liza

tio

n e

rro

r (a

ve

rag

e(m

))

Urban CanyonResidential

Figure 12: Ranging correction effecton AP localization error.

(a) Scalability Exper.

0 8 16 24 32 40 48 56 64 72 80 88 96100

Distance to precise GPS (m)

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

CD

F

Adaptive spb

10spb

2spb

(b) CDF

Figure 13: Evaluating our adaptive FTM message rate ap-proach leveraging FTM ranges only.

Vehicle location error vs. AP localization error. As shownin Fig. 7(a) and Table 2, we notice that the vehicle localization erroris slightly lower than the AP localization error, which is attributedto two factors: first, the multilateration reference points (i.e., vehicletrace) used to locate the AP are collinear, while the reference pointsused to locate the vehicle (i.e., widely spread AP locations) aremostly not, leading to lower dilution of multilateration precision.Second, AP localization is only limited by FTM measurements; theconstraints on vehicle location, on the other hand, come downto GPS, odometry, and FTM measurements. This results in moreaccurate location estimation.

Urban canyon results vs. residential results. The results us-ing the urban canyon and residential datasets consistently showthat FTMSLAM alleviates both GPS blockage in obstructed envi-ronments and occasional GPS noise in unobstructed environments.Interestingly, the average AP localization error is slightly higher inresidential areas, due to signal multipath. We observe that it is im-perative to initialize our system with sufficiently accurate locationas initial conditions, which is possible using GPS in the residentialarea, or using Maps/FTMs in urban canyons. Wi-Go is designed tofilter noisy measurements when it happens, thus only benefitingfrom reasonably accurate measurements. As shown in Fig. 11, thedifferent components of our system have different impacts betweenthe two datasets.

8 RELATEDWORKOutdoor Localization. In urban canyons, GPS can achieve on av-erage 24.3m error [34], which can be reduced by 6-8m as shownrecently [38]. ParkLoc [20] localizes cars with an accuracy of 4.8min underground parking garages using inertial sensors in the smart-phones of the people inside the car, and semi-supervised learningalgorithms. Carloc [34] can track vehicles outdoors (2.7 m meanerror) by matching digital maps and leveraging proprietary vehi-cle sensors to detect roadway landmarks (e.g. speed bumps, stop

signs). The LTE standard [18] indicates that a localization accuracyof 20-30m can be achieved with trilateration methods and usingneural networks in tandem allows for a minimum error of 2.3m tobe possible [50].

WiFi Localization. Indoor localization either assumes theknowledge of APs locations [42, 54] or leverage fingerprintingtechniques [36, 48, 56]. Except for fingerprinting, the position ofthe AP is usually assumed to be known. Locating APs while drivingby throughware-driving could lead to up to 32m error [35]. This canbe improved by estimating angle of arrival reaching 10-30m local-ization error [52]. Fingerprinting strategies, however, involve heavyoverheads, relatively low accuracy compared to ToA methods simi-lar to GPS. CUPID [49] leverages CSI for indoor tracking, achieving2.7m median error. WOLoc [58] offers WiFi RSSI-based outdoortracking through semi-supervised manifold learning technique.

Scalability. Banin et al. [11] proposes a solution using multi-ple access points which are custom made for using WiFi FTM andcommunicating with each other to form a geosynchronous system,where the clients track their positions passively, with only the APstransmitting. Llombart et al. in [39] simulate a technique that canbe used for mobile devices to select three nearby APs for trilater-ation. The mobile device to be positioned finds three nearby APs,associates with each of the three APs, calculates the distance andthen dissociates. Dedicated hardware is required for both the APand the mobile device in this method.

9 CONCLUSIONThis paper introducesWi-Go, a vehicle localization system that usesWiFi FTM measurements to achieve lane-level accuracy in chal-lenging urban canyons where GPS accuracy is degraded. It fusesWiFi FTMs with GPS and odometry in the FTMSLAM framework,to simultaneously estimate vehicle positions and the positions sur-rounding WiFi APs. We further design Wi-Go to adapt the rate ofFTM packets in order to mitigate network congestion and latencyarising from WiFi FTM’s active ranging approach, while maximiz-ing vehicle localization accuracy. We evaluate the system in theurban canyons of Manhattan as well as suburban residential areas.Wi-Go achieves 1.3mmedian error in an urban canyon setting withaccess points on parked vehicles, 2.1m median error in a crowdedarea with access points in buildings, and 0.8 m median localiza-tion error in a suburban environment with access points insideapartment buildings. This shows promise for using WiFi FTM mea-surements for vehicle positioning even in multi-path rich urbancanyons.


REFERENCES[1] https://goo.gl/AW4Pto. Self-Driving Cars Market Growth.[2] https://waymo.com/. Google WAYMO Self-Driving Car.[3] https://goo.gl/BSUCdG. Wi-Fi CERTIFIED Location.[4] https://goo.gl/1z1LhE. Android Pie Indoor Positioning.[5] https://sumo.dlr.de/. Simulation of Urban MObility (SUMO).[6] https://www.openstreetmap.org/. OpenStreetMap.[7] https://www.campllc.org/. Crash Avoidance Metrics Partners LLC (CAMP).[8] "IEEE Standard for Information technology–Telecommunications and informa-

tion exchange between systems Local and metropolitan area networks–Specificrequirements - Part 11: Wireless LAN Medium Access Control (MAC) and Phys-ical Layer (PHY) Specifications". "IEEE Std 802.11-2016 (Revision of IEEE Std802.11-2012)", pages 1–3534, Dec 2016.

[9] H. Aly and M. Youssef. Dejavu: an accurate energy-efficient outdoor localizationsystem. In Proceedings of the 21st ACM SIGSPATIAL International Conference onAdvances in Geographic Information Systems, pages 154–163. ACM, 2013.

[10] P. Bahl, V. N. Padmanabhan, et al. Radar: An in-building rf-based user locationand tracking system.

[11] L. Banin, O. Bar-Shalom, N. Dvorecki, and Y. Amizur. Scalable wi-fi client self-positioning using cooperative ftm-sensors. IEEE Transactions on Instrumentationand Measurement, 2018.

[12] G. Bansal, B. Cheng, A. Rostami, K. Sjoberg, J. B. Kenney, and M. Gruteser.Comparing limeric and dcc approaches for vanet channel congestion control. In2014 IEEE 6th International Symposium on Wireless Vehicular Communications(WiVeC 2014), pages 1–7. IEEE, 2014.

[13] G. Bansal, J. B. Kenney, and C. E. Rohrs. Limeric: A linear adaptive message ratealgorithm for dsrc congestion control. IEEE Transactions on Vehicular Technology,62(9):4182–4197, 2013.

[14] G. Bansal, H. Lu, J. B. Kenney, and C. Poellabauer. Embarc: Error model basedadaptive rate control for vehicle-to-vehicle communications. In Proceeding ofthe tenth ACM international workshop on Vehicular inter-networking, systems, andapplications, pages 41–50. ACM, 2013.

[15] R. Berkvens, A. Jacobson, M. Milford, H. Peremans, and M. Weyn. Biologicallyinspired slam using wi-fi. In 2014 IEEE/RSJ International Conference on IntelligentRobots and Systems, pages 1804–1811. IEEE, 2014.

[16] J.-L. Blanco, J. González, and J.-A. Fernández-Madrigal. A pure probabilisticapproach to range-only slam. In 2008 IEEE International Conference on Roboticsand Automation, pages 1436–1441. IEEE, 2008.

[17] C. Bo, X.-Y. Li, T. Jung, X. Mao, Y. Tao, and L. Yao. Smartloc: Push the limit of theinertial sensor based metropolitan localization using smartphone. In Proceedingsof the 19th annual international conference on Mobile computing & networking,pages 195–198. ACM, 2013.

[18] S. Chairman. Summary of ad-hoc session on study on indoor positioning en-hancements for utra and lte. In 3GPP TSG RAN WG1 Meeting, pages 20–24.

[19] Y.-C. Cheng, Y. Chawathe, A. LaMarca, and J. Krumm. Accuracy characterizationfor metropolitan-scale wi-fi localization. In Proceedings of the 3rd internationalconference on Mobile systems, applications, and services, pages 233–245. ACM,2005.

[20] J. Cherian, J. Luo, and S.-S. Ho. Parkloc: Light-weight graph-based vehicularlocalization in parking garages. Proceedings of the ACM on Interactive, Mobile,Wearable and Ubiquitous Technologies, 2(3):99, 2018.

[21] J. Djugash, S. Singh, and P. Corke. Further results with localization and mappingusing range from radio. In Field and Service Robotics, pages 231–242. Springer,2006.

[22] J. Djugash, S. Singh, G. Kantor, and W. Zhang. Range-only slam for robots oper-ating cooperatively with sensor networks. In Proceedings 2006 IEEE InternationalConference on Robotics and Automation, 2006. ICRA 2006., pages 2078–2084. IEEE,2006.

[23] R. Faragher, C. Sarno, and M. Newman. Opportunistic radio slam for indoornavigation using smartphone sensors. In Proceedings of the 2012 IEEE/ION Position,Location and Navigation Symposium, pages 120–128. IEEE, 2012.

[24] B. Ferris, D. Fox, and N. D. Lawrence. Wifi-slam using gaussian process latentvariable models. In IJCAI, volume 7, pages 2480–2485, 2007.

[25] D. Giustiniano and S. Mangold. Caesar: carrier sense-based ranging in off-the-shelf 802.11 wireless lan. In Proceedings of the Seventh COnference on emergingNetworking EXperiments and Technologies, page 10. ACM, 2011.

[26] M. A. Goudarzi and R. J. Landry. Assessing horizontal positional accuracy ofgoogle earth imagery in the city of montreal, canada. Geodesy and Cartography,43(2):56–65, 2017.

[27] D. Hahnel, W. Burgard, D. Fox, K. Fishkin, and M. Philipose. Mapping andlocalization with rfid technology. In IEEE International Conference on Roboticsand Automation, 2004. Proceedings. ICRA’04. 2004, volume 1, pages 1015–1020.IEEE, 2004.

[28] D. Halperin, W. Hu, A. Sheth, and D. Wetherall. Tool release: Gathering 802.11ntraces with channel state information. ACM SIGCOMM CCR, 41(1):53, Jan. 2011.

[29] S. He and S. . G. Chan. Wi-fi fingerprint-based indoor positioning: Recentadvances and comparisons. IEEE Communications Surveys Tutorials, 18(1):466–490, 2016.

[30] J. Huang, D. Millman, M. Quigley, D. Stavens, S. Thrun, and A. Aggarwal. Efficient,generalized indoor wifi graphslam. In 2011 IEEE International Conference onRobotics and Automation, pages 1038–1043. IEEE, 2011.

[31] M. Ibrahim, H. Liu, M. Jawahar, V. Nguyen, M. Gruteser, R. Howard, B. Yu, andF. Bai. Verification: Accuracy evaluation of wifi fine time measurements on anopen platform. In Proceedings of the 24th Annual International Conference onMobile Computing and Networking, pages 417–427. ACM, 2018.

[32] M. Ibrahim and M. Youssef. Cellsense: An accurate energy-efficient gsm posi-tioning system. IEEE Transactions on Vehicular Technology, 61(1):286–296, 2012.

[33] Y. Jiang, H. Qiu, M. McCartney, W. G. Halfond, F. Bai, D. Grimm, and R. Govindan.Carlog: A platform for flexible and efficient automotive sensing. In Proceedingsof the 12th ACM Conference on Embedded Network Sensor Systems, pages 221–235.ACM, 2014.

[34] Y. Jiang, H. Qiu, M. McCartney, G. Sukhatme, M. Gruteser, F. Bai, D. Grimm, andR. Govindan. Carloc: Precise positioning of automobiles. In Proceedings of the13th ACM Conference on Embedded Networked Sensor Systems, pages 253–265.ACM, 2015.

[35] M. Kim, J. J. Fielding, and D. Kotz. Risks of using ap locations discovered throughwar driving. In International Conference on Pervasive Computing, pages 67–82.Springer, 2006.

[36] A. LaMarca, Y. Chawathe, S. Consolvo, J. Hightower, I. Smith, J. Scott, T. Sohn,J. Howard, J. Hughes, F. Potter, et al. Place lab: Device positioning using radiobeacons in the wild. In International Conference on Pervasive Computing, pages116–133. Springer, 2005.

[37] R. B. Langley et al. Dilution of precision. GPS world, 10(5):52–59, 1999.[38] X. Liu, S. Nath, and R. Govindan. Gnome: A practical approach to nlos mitigation

for gps positioning in smartphones. In Proceedings of the 16th Annual InternationalConference on Mobile Systems, Applications, and Services, pages 163–177. ACM,2018.

[39] M. Llombart, M. Ciurana, and F. Barcelo-Arroyo. On the scalability of a novelwlan positioning system based on time of arrival measurements. In 2008 5thWorkshop on Positioning, Navigation and Communication, pages 15–21. IEEE,2008.

[40] G. Lui, T. Gallagher, B. Li, A. G. Dempster, and C. Rizos. Differences in rssireadings made by different wi-fi chipsets: A limitation of wlan localization. In2011 International conference on localization and GNSS (ICL-GNSS), pages 53–57.IEEE, 2011.

[41] T. Mangel, O. Klemp, and H. Hartenstein. 5.9 ghz inter-vehicle communication atintersections: a validated non-line-of-sight path-loss and fading model. EURASIPJournal on Wireless Communications and Networking, 2011(1):182, 2011.

[42] A. T. Mariakakis, S. Sen, J. Lee, and K.-H. Kim. Sail: Single access point-basedindoor localization. In Proceedings of the 12th annual international conference onMobile systems, applications, and services, pages 315–328. ACM, 2014.

[43] P. Moutarlier and R. Chatila. An experimental system for incremental environ-ment modelling by an autonomous mobile robot. In Experimental Robotics I,pages 327–346. Springer, 1990.

[44] P. Newman and J. Leonard. Pure range-only sub-sea slam. In 2003 IEEE Inter-national Conference on Robotics and Automation (Cat. No. 03CH37422), volume 2,pages 1921–1926. Ieee, 2003.

[45] E. Olson, J. J. Leonard, and S. Teller. Robust range-only beacon localization. IEEEJournal of Oceanic Engineering, 31(4):949–958, 2006.

[46] C. Pei, Z. Wang, Y. Zhao, Z. Wang, Y. Meng, D. Pei, Y. Peng, W. Tang, and X. Qu.Why it takes so long to connect to a wifi access point. In IEEE INFOCOM 2017 -IEEE Conference on Computer Communications, pages 1–9, May 2017.

[47] G. Pulighe, V. Baiocchi, and F. Lupia. Horizontal accuracy assessment of very highresolution google earth images in the city of rome, italy. International Journal ofDigital Earth, 9(4):342–362, 2016.

[48] J. Rekimoto. Placeengine: A wifi location platform based on realworld-folksonomy. In Internet Conference 2006, pages 95–104, 2006.

[49] S. Sen, J. Lee, K.-H. Kim, and P. Congdon. Avoiding multipath to revive inbuildingwifi localization. In Proceeding of the 11th annual international conference onMobile systems, applications, and services, pages 249–262. ACM, 2013.

[50] Z. Shi and Y. Wang. Neural network based localization using outdoor lte mea-surements. In 2018 10th International Conference on Wireless Communicationsand Signal Processing (WCSP), pages 1–6. IEEE, 2018.

[51] R. C. Smith and P. Cheeseman. On the representation and estimation of spatialuncertainty. The international journal of Robotics Research, 5(4):56–68, 1986.

[52] A. P. Subramanian, P. Deshpande, J. Gao, and S. R. Das. Drive-by localization ofroadside wifi networks. In IEEE INFOCOM 2008-The 27th Conference on ComputerCommunications, pages 718–725. IEEE, 2008.

[53] S. Thrun, M. Montemerlo, D. Koller, B. Wegbreit, J. Nieto, and E. Nebot. Fastslam:An efficient solution to the simultaneous localization and mapping problem withunknown data association. Journal of Machine Learning Research, 4(3):380–407,2004.


[54] D. Vasisht, S. Kumar, and D. Katabi. Decimeter-level localization with a singlewifi access point. In 13th {USENIX} Symposium on Networked Systems Designand Implementation ({NSDI} 16), pages 165–178, 2016.

[55] T. Vincenty. Direct and inverse solutions of geodesics on the ellipsoid withapplication of nested equations. Survey review, 23(176):88–93, 1975.

[56] Q. D. Vo and P. De. A survey of fingerprint-based outdoor localization. IEEECommunications Surveys & Tutorials, 18(1):491–506, 2016.

[57] H. Wang, S. Sen, A. Elgohary, M. Farid, M. Youssef, and R. R. Choudhury. Noneed to war-drive: Unsupervised indoor localization. In Proceedings of the 10th

international conference on Mobile systems, applications, and services, pages 197–210. ACM, 2012.

[58] J. Wang, N. Tan, J. Luo, and S. J. Pan. Woloc: Wifi-only outdoor localizationusing crowdsensed hotspot labels. In IEEE INFOCOM 2017-IEEE Conference onComputer Communications, pages 1–9. IEEE, 2017.

[59] J. Xiong and K. Jamieson. Arraytrack: A fine-grained indoor location system. InPresented as part of the 10th {USENIX} Symposium on Networked Systems Designand Implementation ({NSDI} 13), pages 71–84, 2013.

wi-go: accurate and scalable vehicle positioning using wifi fine...

Documents